Method for constructing power grid physical asset evaluation analysis model under multi-boundary condition

Abstract

The invention discloses a power grid physical asset evaluation analysis model construction method under a multi-boundary condition. The method comprises the following steps: S1, acquiring multi-dimensional historical boundary condition data of power grid physical assets; s2, performing normalization processing on the multi-dimensional historical boundary condition data to generate standardized historical boundary condition data; s3, constructing an adaptive evaluation model based on the dimension type of the standardized historical boundary condition data; s4, training the adaptive evaluation model, and optimizing parameters through a machine learning algorithm to obtain a trained adaptive evaluation model; and S5, obtaining multi-dimensional real-time boundary condition data of the current power grid physical assets, performing normalization processing, generating standardized real-time boundary condition data, inputting the standardized real-time boundary condition data into the trained adaptive evaluation model, and outputting an evaluation result of the current power grid physical assets. The power grid physical asset evaluation accuracy and reliability are high, the adaptability to the complex dynamic environment is good, and the fine management requirements of power grid enterprises are met.

Description

Technical Field

[0001] This invention relates to the field of power grid physical asset evaluation and analysis technology, specifically to a method for constructing a power grid physical asset evaluation and analysis model under multiple boundary conditions. Background Technology

[0002] With the continuous expansion of the power grid and the increasing complexity of its operating environment, the evaluation and analysis of power grid physical assets faces numerous challenges. While various existing models exist for evaluating and analyzing power grid physical assets, most are optimized for specific boundary conditions and lack the ability to dynamically adapt to multiple boundary conditions. In practical applications, the evaluation and analysis of power grid physical assets needs to comprehensively consider multiple boundary conditions, such as equipment operating status, environmental factors, and power grid topology. Existing models often exhibit limitations when handling these complex and variable boundary conditions, affecting the accuracy and reliability of the evaluation results and making it difficult to meet the needs of power grid companies for refined management of physical assets. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the above-mentioned defects of the prior art and provide a method for constructing a power grid physical asset evaluation and analysis model under multiple boundary conditions that has high accuracy and reliability in power grid physical asset evaluation, good adaptability to complex dynamic environments, and meets the needs of power grid enterprises for refined management.

[0004] The technical solution adopted by this invention to solve its technical problem is as follows: A method for constructing a power grid physical asset evaluation and analysis model under multiple boundary conditions, comprising the following steps: S1. Obtain multi-dimensional historical boundary condition data of power grid physical assets; the multi-dimensional historical boundary condition data includes historical equipment operating status data, historical environmental factor data, and historical power grid topology data. S2. Normalize the multi-dimensional historical boundary condition data to generate standardized historical boundary condition data. S3. Based on the dimension type of the standardized historical boundary condition data, construct an adaptive evaluation model; the adaptive evaluation model includes multiple sub-models and a dynamic weight allocation module, wherein each sub-model is configured to process boundary condition data of a specific dimension, and the dynamic weight allocation module is configured to calculate the weight coefficients of each sub-model in real time according to the characteristics of the input data and perform weighted fusion. S4. Using the historical dataset composed of the standardized historical boundary condition data and corresponding evaluation labels, the adaptive evaluation model is trained, and the parameters of the adaptive evaluation model are optimized by machine learning algorithms to obtain the trained adaptive evaluation model. S5. Obtain multi-dimensional real-time boundary condition data of the current power grid physical assets, and normalize the multi-dimensional real-time boundary condition data to generate standardized real-time boundary condition data; input the standardized real-time boundary condition data into the trained adaptive evaluation model, and output the evaluation result of the current power grid physical assets; the evaluation result includes asset health status score and risk warning information.

[0005] Preferably, in S1, the specific steps include: The historical equipment operating status data is obtained from the power grid production management system, the historical environmental factor data is obtained from the environmental monitoring platform, and the historical power grid topology data is obtained from the power grid graphic management platform. The historical equipment operating status data, the historical environmental factor data, and the historical power grid topology data are timestamped and missing values ​​are processed to obtain the multi-dimensional historical boundary condition data with time consistency.

[0006] Preferably, in step S2, the specific steps include: For the historical equipment operating status data, the historical environmental factor data, and the historical power grid topology data after timestamp alignment and missing value processing, the normalization parameter range corresponding to each dimension of data is determined respectively. The data for each dimension is linearly transformed to a uniform preset numerical range using either the extreme value method or the standard deviation standardization method. Based on the numerical distribution characteristics of the data for each dimension, an appropriate transformation algorithm is selected to generate the standardized historical boundary condition data with uniform dimensions.

[0007] Preferably, in S3, the specific steps for constructing the adaptive evaluation model include: S31. Based on the three dimensions of historical equipment operating status data, historical environmental factor data, and historical power grid topology data in the standardized historical boundary condition data, configure three independent sub-models accordingly. S32. The dynamic weight allocation module is constructed as a learnable parameter network, which takes the standardized boundary condition data input in real time as its feature and outputs dynamic weight coefficients corresponding to the three sub-models. S33. The outputs of the three sub-models are weighted and summed with the dynamic weight coefficients calculated by the dynamic weight allocation module to construct the adaptive evaluation model.

[0008] Preferably, in S31, the sub-model for processing historical equipment operating status data adopts a gated recurrent unit network structure to extract time-dependent features from the equipment operating status data; the sub-model for processing historical environmental factor data adopts a fully connected neural network structure to learn the nonlinear relationship between environmental factors and equipment status; and the sub-model for processing historical power grid topology data adopts a graph neural network structure to capture the topological connection relationship between power grid nodes.

[0009] Preferably, S33 further includes: S331. The output of the sub-model corresponding to the historical equipment operating status data is denoted as f. S (X S The output of the sub-model corresponding to the historical environmental factor data is denoted as f. E (X E The output of the sub-model corresponding to the historical power grid topology data is denoted as f. T (X T ); S332. The dynamic weight allocation module calculates the dynamic weight coefficients corresponding to the three sub-models based on the real-time input standardized boundary condition data X, denoted as w respectively. S (X), w E (X), w T (X), and satisfy w S (X)+w E (X)+w T (X)=1; S333. Integrate the outputs of the three sub-models with the dynamic weight coefficients to generate the evaluation result Y, so as to construct the adaptive evaluation model that represents the standardized boundary condition data X and the evaluation output Y.

[0010] Preferably, in step S333, the outputs of the three sub-models are integrated with the dynamic weight coefficients using a preset fusion formula to generate the evaluation result Y; the preset fusion formula is: Y = w S (X)* f S (X S )+ w E (X) * f E (X E )+w T (X) * f T (X T ); Where X represents the standardized boundary condition data input in real time, X S X E X Tf represents the feature vectors extracted from X corresponding to equipment operating status, environmental factors, and power grid topology, respectively. S (·), f E (·), f T (·) represent the mapping functions corresponding to the three sub-models, w S (X), w E (X), w T (X) represents the dynamic weight coefficient generated by the dynamic weight allocation module, which depends on the real-time input X, and Y represents the evaluation result calculated by the adaptive evaluation model.

[0011] Preferably, in step S4, the specific steps include: Initialize the parameters of the adaptive evaluation model, including the internal parameters of the three sub-models and the learnable parameters of the dynamic weight allocation module; The historical dataset is divided into a training set and a validation set; The gradient descent algorithm is used to minimize the loss function between the model output and the evaluation label. Multiple rounds of iterative training are performed on the training set, and the model performance is monitored on the validation set to prevent overfitting. Training stops when the model's performance on the validation set reaches a preset standard. The parameter state of the model at this point is saved to obtain the trained adaptive evaluation model.

[0012] Preferably, in step S5, the specific steps for input and output include: S51. Input the standardized real-time boundary condition data into the trained adaptive evaluation model, and calculate the dynamic weight coefficients under the current real-time conditions through the dynamic weight allocation module. S52. Input the standardized real-time boundary condition data vector X into the trained adaptive evaluation model, and calculate the dynamic weight coefficient w under the current real-time conditions through the dynamic weight allocation module. S (X), w E (X), w T (X), and simultaneously calculate f through three sub-models respectively. S (X S ), f E (X E ), f T (X T ); S53. Calculate the asset health status score according to the preset fusion formula, compare the asset health status score with the preset risk level threshold, generate corresponding risk warning information based on the comparison result, and output the evaluation result containing the asset health status score and the risk warning information.

[0013] Preferably, in S53, the specific steps of the comparison include: Multiple risk level thresholds are set to divide the scoring range into four levels: normal state, attention state, abnormal state, and emergency state. When the score falls into the attention state or a higher risk level, a risk warning message containing a specific risk level description and suggested handling measures is generated. The risk level thresholds are dynamically adjusted according to the probability distribution of historical evaluation results.

[0014] The beneficial effects of the model construction method of the present invention are as follows: By acquiring and normalizing multi-dimensional historical boundary condition data, and combining it with a dynamic weight allocation module to adjust the weights of each sub-model in real time, the model construction method of the present invention optimizes the model parameters using machine learning algorithms, thereby improving the accuracy and reliability of power grid physical asset evaluation, solving the problem of insufficient adaptability of existing models to complex dynamic environments, and meeting the refined management needs of power grid enterprises. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an embodiment of the method for constructing a power grid physical asset evaluation and analysis model under multiple boundary conditions according to the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0017] Example: like Figure 1 As shown, this invention provides a method for constructing a power grid physical asset evaluation and analysis model under multiple boundary conditions, executed by a control terminal, and includes the following steps: S1. Obtain multi-dimensional historical boundary condition data of power grid physical assets; the multi-dimensional historical boundary condition data includes historical equipment operating status data, historical environmental factor data, and historical power grid topology data. Among them, "power grid physical assets" refers to fixed assets with physical form within a power grid enterprise; for example, a 220kV transformer belonging to a provincial power grid company. "Multi-dimensional historical boundary condition data" refers to a collection of multi-source data reflecting the operational status of power grid physical assets collected over past periods; for example, load rate, ambient temperature, and wiring data of the 220kV transformer recorded from January to December 2025. "Historical equipment operating status data" refers to parameter data characterizing the historical operating performance of power grid physical assets; for example, oil temperature, winding temperature, and dissolved gas content data of the 220kV transformer in 2025. "Historical environmental factor data" refers to historical data on the external natural environment affecting the operation of power grid physical assets; for example, ambient temperature, humidity, and wind speed data of the 220kV transformer's installation location in 2025. "Historical power grid topology data" refers to structured data describing the historical connection relationships of the power grid; for example, the main electrical wiring diagram and circuit breaker status information of the substation where the 220kV transformer is located in 2025. S2. Normalize the multi-dimensional historical boundary condition data to generate standardized historical boundary condition data. Standardized historical boundary condition data refers to historical data that has undergone normalization to eliminate dimensional differences; for example, the oil temperature (unit: °C) and load rate (unit: %) of the 220kV transformer are uniformly transformed to values ​​within the range of [0,1]. S3. Based on the dimension type of the standardized historical boundary condition data, construct an adaptive evaluation model; the adaptive evaluation model includes multiple sub-models and a dynamic weight allocation module, wherein each sub-model is configured to process boundary condition data of a specific dimension, and the dynamic weight allocation module is configured to calculate the weight coefficients of each sub-model in real time according to the characteristics of the input data and perform weighted fusion. In this context, "dimensional type" refers to the category of boundary condition data based on its physical meaning and source; for example, data can be divided into three dimensional types: equipment operating status, environmental factors, and power grid topology. "Adaptive evaluation model" refers to an evaluation model that can dynamically adjust its internal parameters based on input conditions; for example, a machine learning model used to assess the health status of the 220kV transformer. "Sub-model" refers to an independent functional module within the adaptive evaluation model that processes data of a specific dimension; for example, a GRU neural network module specifically analyzing the oil temperature variation trend of the 220kV transformer. "Dynamic weight allocation module" refers to a component that automatically calculates the contribution weights of each sub-model based on real-time input features; for example, a fully connected neural network that takes real-time data as input and outputs three weight values. "Specific dimension boundary condition data" refers to boundary condition data belonging to a certain dimensional type; for example, specifically referring to the environmental temperature data of the 220kV transformer. "Weight coefficient" refers to a numerical value representing the importance of the corresponding sub-model's output in the final evaluation; for example, the weight of the environmental factor sub-model under the current real-time conditions is calculated to be 0.3. S4. Using the historical dataset composed of the standardized historical boundary condition data and corresponding evaluation labels, the adaptive evaluation model is trained, and the parameters of the adaptive evaluation model are optimized by machine learning algorithms to obtain the trained adaptive evaluation model. The evaluation label refers to the actual asset evaluation value in the historical dataset; for example, an expert might label the health status of the 220kV transformer on a certain day in 2025 as "good" (corresponding to a score of 90). The historical dataset refers to the collection of standardized historical boundary condition data and their evaluation labels used for model training; for example, a data table containing standardized data for the entire year of 2025 and corresponding scores. The machine learning algorithm refers to the optimization algorithm used to learn model parameters from the data; for example, using the Adam optimizer to train the adaptive evaluation model. The trained adaptive evaluation model refers to a model that has completed parameter optimization and can be used for actual prediction; for example, a model that has been trained and can receive real-time data and output the health score of the 220kV transformer. S5. Obtain multi-dimensional real-time boundary condition data of the current power grid physical assets, and normalize the multi-dimensional real-time boundary condition data to generate standardized real-time boundary condition data; input the standardized real-time boundary condition data into the trained adaptive evaluation model, and output the evaluation result of the current power grid physical assets; the evaluation result includes asset health status score and risk warning information; The current physical assets of the power grid refer to the specific power grid asset to be evaluated; for example, the 220kV transformer whose health status needs to be assessed. Multi-dimensional real-time boundary condition data refers to boundary condition data collected from multiple sources at the current moment; for example, the real-time oil temperature, ambient temperature, and circuit breaker status of the 220kV transformer monitored at the current moment. Standardized real-time boundary condition data refers to real-time data after undergoing the same normalization process as historical data; for example, transforming the current oil temperature reading to a value within the [0,1] range. Evaluation results refer to the model's calculation output of the current asset status; for example, the health score and risk warning calculated by the model for the 220kV transformer. Asset health status score refers to the numerical result that quantifies the degree of asset health; for example, the model outputs a current health score of 85 for the 220kV transformer. Risk warning information refers to text warning information generated based on the score results; for example, when the score is 85, the system generates a prompt: "Pay attention to the status, it is recommended to strengthen inspections."

[0018] This invention improves the accuracy and reliability of power grid physical asset evaluation by acquiring and normalizing multi-dimensional historical boundary condition data, adjusting the weights of each sub-model in real time using a dynamic weight allocation module, and optimizing model parameters using machine learning algorithms. This solves the problem of insufficient adaptability of existing models to complex dynamic environments and meets the refined management needs of power grid enterprises.

[0019] In one alternative approach, S1 includes the following specific steps: The historical equipment operating status data is obtained from the power grid production management system, the historical environmental factor data is obtained from the environmental monitoring platform, and the historical power grid topology data is obtained from the power grid graphic management platform. The power grid production management system refers to a business system that manages the operation data of power grid equipment; for example, obtaining historical defect records of the 220kV transformer from the PMS system. The environmental monitoring platform refers to a data platform that collects environmental parameters; for example, obtaining historical temperature data of the transformer installation location from a meteorological monitoring platform. The power grid graphic management platform refers to a system that manages power grid topology graphic data; for example, obtaining the electrical wiring diagram of the substation where the transformer is located from the GPMS platform. The historical equipment operating status data, the historical environmental factor data, and the historical power grid topology data are timestamped and missing values ​​are processed to obtain the multi-dimensional historical boundary condition data with time consistency.

[0020] Among the above-mentioned optional methods, the accurate fusion of multi-source data is further achieved. Through timestamp alignment and missing value handling, the historical boundary condition data is ensured to have temporal consistency, providing a reliable data foundation for subsequent model training and asset evaluation.

[0021] In one alternative approach, in S2, the specific steps include: For the historical equipment operating status data, the historical environmental factor data, and the historical power grid topology data after timestamp alignment and missing value processing, the normalization parameter range corresponding to each dimension of data is determined respectively. The data for each dimension is linearly transformed to a uniform preset numerical range using either the extreme value method or the standard deviation standardization method. Based on the numerical distribution characteristics of the data for each dimension, an appropriate transformation algorithm is selected to generate the standardized historical boundary condition data with uniform dimensions.

[0022] Among the above-mentioned optional methods, a normalization process is further introduced to eliminate the dimensional differences of data in different dimensions. The data is unified to a preset range by using extreme value method or standard deviation standardization method, thereby optimizing the model input features and improving the model's adaptability and processing efficiency for complex data.

[0023] In an alternative approach, in S3, the specific steps for constructing the adaptive evaluation model include: S31. Based on the three dimensions of historical equipment operating status data, historical environmental factor data, and historical power grid topology data in the standardized historical boundary condition data, configure three independent sub-models accordingly. S32. The dynamic weight allocation module is constructed as a learnable parameter network, which takes the standardized boundary condition data input in real time as its feature and outputs dynamic weight coefficients corresponding to the three sub-models. Learnable parameter networks refer to neural networks whose parameters can be optimized through training data; for example, the weight matrix in a dynamic weight allocation module that can be adjusted by gradient descent. S33. The outputs of the three sub-models are weighted and summed with the dynamic weight coefficients calculated by the dynamic weight allocation module to construct the adaptive evaluation model.

[0024] Among the above-mentioned optional approaches, an adaptive evaluation model architecture is further designed, with three independent sub-models configured to focus on processing equipment, environment, and topology data respectively. These are paired with a dynamic weight allocation module with learnable parameters to calculate the weights of each sub-model in real time and accurately, thereby achieving intelligent fusion and evaluation of multi-dimensional data.

[0025] In an optional manner, in S31, the sub-model for processing historical equipment operating status data adopts a gated recurrent unit network structure to extract time-dependent features from the equipment operating status data; the sub-model for processing historical environmental factor data adopts a fully connected neural network structure to learn the nonlinear relationship between environmental factors and equipment status; and the sub-model for processing historical power grid topology data adopts a graph neural network structure to capture the topological connection relationships between power grid nodes. Among them, the gated recurrent unit network structure refers to a recurrent neural network structure used for processing sequential data; for example, a GRU network used to analyze the time series data of 220kV transformer oil temperature. The fully connected neural network structure refers to a neural network structure where every neuron in each layer is fully connected to the next layer; for example, a three-layer perceptron used to learn the nonlinear relationship between ambient temperature and transformer lifespan. The graph neural network structure refers to a neural network structure used for processing graph-structured data; for example, a GNN model used to analyze the electrical connection relationships in substation wiring diagrams.

[0026] Among the above-mentioned optional methods, the model architecture is further improved to adapt to multi-dimensional data characteristics. Gated recurrent unit networks are used to extract the time-dependent features of equipment operating status, fully connected neural networks are used to learn the nonlinear relationship between environmental factors and equipment status, and graph neural networks are used to capture the node connection characteristics of the power grid topology, thereby comprehensively improving the model's ability to analyze complex boundary conditions of power grid assets.

[0027] In an alternative embodiment, S33 further includes: S331. The output of the sub-model corresponding to the historical equipment operating status data is denoted as f. S (X S The output of the sub-model corresponding to the historical environmental factor data is denoted as f. E (X E The output of the sub-model corresponding to the historical power grid topology data is denoted as f. T (X T ); S332. The dynamic weight allocation module calculates the dynamic weight coefficients corresponding to the three sub-models based on the real-time input standardized boundary condition data X, denoted as w respectively. S (X), w E (X), w T (X), and satisfy w S (X)+w E (X)+w T (X)=1; S333. Integrate the outputs of the three sub-models with the dynamic weight coefficients to generate the evaluation result Y, so as to construct the adaptive evaluation model that represents the standardized boundary condition data X and the evaluation output Y.

[0028] Among the above-mentioned optional methods, the dynamic weight allocation mechanism is further clarified. Based on the standardized boundary condition data input in real time, the sub-model output is integrated with the dynamic weight coefficient to generate an evaluation result that comprehensively reflects the asset status, thereby enhancing the model's dynamic adaptability to real-time operating conditions and the accuracy of the evaluation.

[0029] In an optional approach, in step S333, the outputs of the three sub-models are integrated with the dynamic weight coefficients using a preset fusion formula to generate the evaluation result Y; the preset fusion formula is: Y = w S (X)* f S (X S )+w E (X) * f E (X E )+w T (X) * f T (X T ); Where X represents the standardized boundary condition data input in real time, X S X E X T f represents the feature vectors extracted from X corresponding to equipment operating status, environmental factors, and power grid topology, respectively. S (·), f E (·), f T (·) represent the mapping functions corresponding to the three sub-models, w S (X), w E (X), w T (X) represents the dynamic weight coefficient generated by the dynamic weight allocation module, which depends on the real-time input X, and Y represents the evaluation result calculated by the adaptive evaluation model.

[0030] In the above-mentioned optional methods, the sub-model output and dynamic weight coefficients are further weighted and fused by a preset fusion formula to achieve dynamic integration and real-time evaluation of multi-dimensional data, thereby improving the overall accuracy and reliability of asset evaluation.

[0031] In one alternative approach, in S4, the specific steps include: Initialize the parameters of the adaptive evaluation model, including the internal parameters of the three sub-models and the learnable parameters of the dynamic weight allocation module; The historical dataset is divided into a training set and a validation set; The training set refers to the subset of data used for learning model parameters; for example, 70% of the data from 2025 is randomly selected for training. The validation set refers to the subset of data used for tuning hyperparameters and monitoring the training process; for example, another 15% of the data from 2025 is selected for validation. The gradient descent algorithm is used to minimize the loss function between the model output and the evaluation label. Multiple rounds of iterative training are performed on the training set, and the model performance is monitored on the validation set to prevent overfitting. The loss function refers to a function that measures the difference between the model's predicted value and the actual value; for example, the mean squared error function is used to calculate the difference between the model's output score and the expert score. Training stops when the model's performance on the validation set reaches a preset standard. The parameter state of the model at this point is saved to obtain the trained adaptive evaluation model.

[0032] Among the above optional methods, further optimization of model parameters is achieved through multiple rounds of iterative training and validation set monitoring, using the gradient descent algorithm to improve the model's adaptability to complex data and prediction accuracy, effectively preventing overfitting.

[0033] In one alternative approach, the specific steps for input and output in S5 include: S51. Input the standardized real-time boundary condition data into the trained adaptive evaluation model, and calculate the dynamic weight coefficients under the current real-time conditions through the dynamic weight allocation module. The current real-time conditions refer to the specific data characteristics at the moment the model is performing inference; for example, at the current moment, the load rate of the 220kV transformer is 75% and the ambient temperature is 25℃. S52. Input the standardized real-time boundary condition data vector X into the trained adaptive evaluation model, and calculate the dynamic weight coefficient w under the current real-time conditions through the dynamic weight allocation module. S (X), w E (X), w T (X), and simultaneously calculate f through three sub-models respectively. S (X S ), f E (X E ), f T (X T ); S53. Calculate the asset health status score according to the preset fusion formula, compare the asset health status score with the preset risk level threshold, generate corresponding risk warning information according to the comparison result, and output the evaluation result containing the asset health status score and the risk warning information. The preset risk level threshold refers to the pre-set scoring boundary used to classify risk levels; for example, a score above 90 is considered normal, 80-90 is considered alert, 70-80 is considered abnormal, and below 70 is considered urgent.

[0034] In the above-mentioned optional methods, standardized real-time data is input into the trained model, and the weight coefficients are calculated in real time with the help of the dynamic weight allocation module. The three sub-models are activated simultaneously, and the asset health status score is quickly output according to the preset formula. Combined with the risk level threshold comparison, early warning information is generated in real time, which further improves the real-time asset evaluation capability and achieves efficient and accurate dynamic assessment.

[0035] In an alternative approach, S53 includes the following specific steps for the comparison: Multiple risk level thresholds are set to divide the scoring range into four levels: normal state, attention state, abnormal state, and emergency state. When the score falls into the attention state or a higher risk level, a risk warning message containing a specific risk level description and suggested handling measures is generated. The risk level thresholds are dynamically adjusted according to the probability distribution of historical evaluation results. The terms "Normal State" and "Accidental State" are as follows: "Normal State" refers to an asset health score within the optimal range; for example, a score above 90 indicates the 220kV transformer is operating well. "Attention State" refers to a slightly declining asset health score requiring attention; for example, a score between 80-90 suggests increased monitoring frequency for the 220kV transformer. "Abnormal State" refers to a significantly declining asset health score requiring intervention; for example, a score between 70-80 suggests potential defects in the 220kV transformer. "Emergency State" refers to an extremely low asset health score requiring immediate action; for example, a score below 70 indicates a high risk and necessitates shutdown and maintenance of the 220kV transformer. "Recommended Actions" refers to specific operational suggestions included in the warning information; for example, for a "Attention State" warning, the recommended action is to "arrange infrared temperature measurement next week." The probability distribution of historical evaluation results refers to the statistical regularity of all evaluation results over a past period; for example, based on 2025 scoring data, 90% of the scores were above 80, and the threshold was adjusted accordingly.

[0036] Among the above-mentioned optional methods, multi-level risk level thresholds are set to dynamically divide the scoring range. When the asset health status score discrepancies to the risk level of attention or above, an early warning message containing risk description and handling suggestions is generated in real time, which further enhances the accuracy and practicality of risk warning and assists in operation and maintenance decision-making. Furthermore, the value of historical data is further explored to optimize risk assessment. The risk level threshold is dynamically adjusted according to the probability distribution of historical evaluation results to generate accurate risk warning information, making the risk warning mechanism more in line with the actual operating conditions of power grid assets and improving the risk prediction accuracy and adaptability of the asset evaluation model.

Claims

1. A method for constructing a power grid physical asset evaluation analysis model under multi-boundary conditions, characterized in that, The method comprises the following steps: S1, obtaining multi-dimensional historical boundary condition data of power grid physical assets; the multi-dimensional historical boundary condition data comprises historical equipment operating state data, historical environmental factor data and historical power grid topology structure data; S2, performing normalization processing on the multi-dimensional historical boundary condition data to generate standardized historical boundary condition data; S3, constructing an adaptive evaluation model based on the dimension types of the standardized historical boundary condition data; the adaptive evaluation model comprises a plurality of sub-models and a dynamic weight distribution module, wherein each sub-model is configured to process boundary condition data of a specific dimension, and the dynamic weight distribution module is configured to calculate weight coefficients of each sub-model in real time according to characteristics of input data and perform weighted fusion; S4, training the adaptive evaluation model by using a historical data set composed of the standardized historical boundary condition data and corresponding evaluation labels, optimizing parameters of the adaptive evaluation model by a machine learning algorithm, and obtaining a trained adaptive evaluation model; S5, obtaining multi-dimensional real-time boundary condition data of current power grid physical assets, performing normalization processing on the multi-dimensional real-time boundary condition data to generate standardized real-time boundary condition data, inputting the standardized real-time boundary condition data into the trained adaptive evaluation model, and outputting an evaluation result of the current power grid physical assets; the evaluation result comprises an asset health state score and risk warning information.

2. The method of claim 1, wherein the method further comprises: In S1, the specific steps comprise: obtaining the historical equipment operating state data from a power grid production management system, obtaining the historical environmental factor data from an environmental monitoring platform, and obtaining the historical power grid topology structure data from a power grid graphics management platform; performing timestamp alignment and missing value processing on the historical equipment operating state data, the historical environmental factor data and the historical power grid topology structure data to obtain the multi-dimensional historical boundary condition data with time consistency.

3. The method of claim 1 or 2, wherein the method further comprises: In S2, the specific steps comprise: determining a normalization parameter range corresponding to each dimension data for the historical equipment operating state data, the historical environmental factor data and the historical power grid topology structure data after timestamp alignment and missing value processing; linearly transforming data of each dimension to a unified preset numerical interval by using an extreme value method or a standard deviation standardization method, and selecting a corresponding transformation algorithm based on numerical distribution characteristics of the data of each dimension to generate the standardized historical boundary condition data with unified dimensions.

4. The method of claim 1 or 2, wherein the method further comprises: In S3, the specific steps of constructing the adaptive evaluation model comprise: S31, configuring three independent sub-models according to the three dimension types of the historical equipment operating state data, the historical environmental factor data and the historical power grid topology structure data in the standardized historical boundary condition data; S32, constructing the dynamic weight distribution module as a learnable parameter network, wherein the learnable parameter network takes real-time input standardized boundary condition data as characteristics and outputs dynamic weight coefficients corresponding to the three sub-models; S33, weighting and summing the outputs of the three sub-models with the dynamic weight coefficients calculated by the dynamic weight distribution module to construct the adaptive evaluation model.

5. The method of claim 4, wherein the method further comprises: In S31, the sub-model for processing historical equipment operating state data adopts a gated recurrent unit network structure to extract time-dependent features in the equipment operating state data. The sub-model for processing historical environmental factor data adopts a fully connected neural network structure to learn the nonlinear relationship between environmental factors and equipment state.

6. The method of claim 4, wherein the method further comprises: In S33, further comprising: S331, record the output of the sub-model corresponding to the historical equipment operation state data as f S (X S ), record the output of the sub-model corresponding to the historical environment factor data as f E (X E ), record the output of the sub-model corresponding to the historical power grid topology data as f T (X T ); S332、through the dynamic weight distribution module according to the real-time input of standardized boundary condition data X, the dynamic weight coefficient corresponding to three sub-models is calculated, respectively recorded as w S (X), w E (X), w T (X), and w S (X)+w E (X)+w T (X)=1; S333, integrating the outputs of the three sub-models with the dynamic weight coefficients to generate an evaluation result Y, so as to construct the adaptive evaluation model representing the standardized boundary condition data X and the evaluation output Y.

7. The method of claim 6, wherein the method further comprises: In S333, the outputs of the three sub-models are integrated with the dynamic weight coefficient by a preset fusion formula to generate the evaluation result Y; the preset fusion formula is: Y = w S (X)* f S (X S )+ w E (X) * f E (X E )+w T (X) *f T (X T ). wherein X represents the real-time input standardized boundary condition data, X S , X E , X T respectively represent the feature vectors corresponding to the equipment operating state, environmental factors and power grid topology extracted from X, f S (·), f E (·), f T (·) respectively represent the mapping functions corresponding to the three sub-models, w S (X), w E (X), w T (X) represent the dynamic weight coefficients generated by the dynamic weight distribution module, which depend on the real-time input X, and Y represents the evaluation result calculated by the adaptive evaluation model.

8. The method of claim 1 or 2, wherein the method further comprises: In S4, the specific steps include: Initializing the parameters of the adaptive evaluation model, including the internal parameters of the three sub-models and the learnable parameters of the dynamic weight distribution module; Dividing the historical data set into a training set and a validation set; Using a gradient descent algorithm to minimize the loss function between the model output and the evaluation label, performing multiple rounds of iterative training on the training set, and monitoring the model performance on the validation set to prevent overfitting; Stopping training when the performance of the model on the validation set reaches the preset standard, saving the parameter state of the model at this time, and obtaining the trained adaptive evaluation model.

9. The method of claim 1 or 2, wherein the method further comprises: In S5, the specific steps of input and output include: S51, inputting the standardized real-time boundary condition data into the trained adaptive evaluation model to calculate the dynamic weight coefficients under the current real-time condition through the dynamic weight distribution module; S52, input the standardized real-time boundary condition data vector X to the trained adaptive evaluation model, and calculate the dynamic weight coefficient w under the current real-time condition through the dynamic weight distribution module S (X), w E (X), w T (X), and three sub-models are used to calculate f S (X S ), f E (X E ), f T (X T ) respectively and synchronously. S53, calculating the asset health state score according to a preset fusion formula, comparing the asset health state score with a preset risk level threshold, and generating corresponding risk warning information according to the comparison result, and outputting the evaluation result containing the asset health state score and the risk warning information.

10. The method of claim 9, wherein the method further comprises: In S53, the specific steps of comparison include: Setting multiple risk level thresholds to divide the score range into four levels of normal state, attention state, abnormal state and emergency state; when the score falls into the attention state or higher risk level, generating risk warning information containing specific risk level description and recommended treatment measures; wherein the risk level threshold is dynamically adjusted according to the probability distribution of historical evaluation results.

Images